UNET NEURAL NETWORK IN AGRICULTURAL LAND COVER CLASSIFICATION USING SENTINEL-2

Abstract

The article discusses a method for classifying land cover types in rural areas using a trained neural network. The focus is on distinguishing agriculturally cultivated areas and differentiating bare soil from quarry areas. This distinction is not present in publicly available databases like CORINE, UrbanAtlas, EuroSAT, or BigEarthNet. The research involves training a neural network on multi-temporal patches to classify Sentinel-2 images rapidly. This approach allows automated monitoring of cultivated areas, determining periods of bare soil vulnerability to erosion, and identifying open-pit areas with similar spectral characteristics to bare soil. After training the U-Net network, it achieved an average classification accuracy of 90% (OA) in the test areas, highlighting the importance of using OA for multi-class classifications, instead of ACC. Analysis of our main classes revealed high accuracy, 99.01% for quarries, 92.3% for bare soil, and an average of 94.8% for annual crops, demonstrating the model's capability to differentiate between crops at various growth stages and assess land cover categories effectively.